ToolsDocsSign In
← All posts

Querying your saved library with Claude (using MCP)

May 2, 2026

You read a Cal Newport interview on YouTube last month. You highlighted half a Kindle book about deep work in February. You bookmarked four essays on attention from X over the weekend. You can feel that all of it is somewhere in your saved library. You cannot remember where, and you do not want to scroll.

This is what an MCP server fixes. You point Claude at your library, ask in plain English, and the matching saves come back, sometimes with the exact passage already pulled out.

Keep is one of the few read-later libraries that ships an MCP server. One config line and Claude Code, Claude Desktop, or Cursor can read across every article, book highlight, YouTube transcript, and X bookmark you have saved.

What MCP actually is, in one sentence

Model Context Protocol is an open standard for connecting AI clients to outside systems. Anthropic shipped it. Claude, ChatGPT, Cursor, and a growing list of other clients all speak it. If a tool publishes an MCP server, any of those clients can use it.

That is the whole reason this works. Keep does not need its own AI panel. Your existing client gets the access.

Real prompts you can run against your own library

These are the kinds of prompts I actually use. The answers below are the real shape Claude returns when the matching saves exist in your library.

"Show me everything I saved about deep work this year"

Claude calls search_items against your library, filters by save date, and lists what it finds. You get titles, sources, and a one-line summary per item, grouped roughly by theme. Articles, books, X threads, and YouTube videos all show up in the same list because Keep stores all of them as markdown.

The useful version of this prompt is the one with a follow-up. "Now pull the three passages most worth re-reading" is when it earns its keep, because Claude can open the full body of any item with get_item and read the whole thing before answering.

"Summarise what I read this week"

A weekly skim on autopilot. Claude lists every item from the last seven days, groups them by source type, and writes a paragraph per cluster. You learn that you saved four pieces about the same topic without realising it. That alone is usually a signal.

This prompt is also the one I run on Sunday mornings before I write anything.

"Find that quote about flow from a Kindle book"

This is the cross-source query the official Kindle notebook viewer cannot do. Your Kindle highlights live alongside your articles in Keep, so Claude searches across all of them. It returns the passage, the book it came from, and a link back to the entry. If you only half-remember the wording, give it the half you remember; semantic search closes most of the gap.

"What did Cal Newport say in the interview I bookmarked on YouTube last month"

Keep saves YouTube videos with their transcript, so Claude reads the transcript, finds the section that matches your question, and quotes it back with a timestamp. You did not need to scrub. You did not need to rewatch. You asked, and the answer came back.

What the server actually exposes

Keep's MCP server exposes 13 tools. The ones you will hit most:

  • search_items for keyword and semantic search across your library.
  • list_items for filtered listing (by source, tag, date range, status).
  • get_item to pull the full markdown body of one save.
  • list_item_highlights and get_highlight for passage-level retrieval.
  • save_item so Claude can add new saves on your behalf.
  • whoami, get_stats, list_inbox, mark_done for account and inbox plumbing.

Claude picks the tool. You write the prompt.

How to install it (one line)

If you use Claude Code:

claude mcp add --transport http keep https://keep.md/mcp

That is the install. Claude Code will walk you through the OAuth flow on first use, and after that your library is readable from any conversation.

In Claude Desktop or Cursor, point the client at the same endpoint (https://keep.md/mcp) using whatever MCP config method that client supports. The transport is HTTP. The auth is OAuth, scoped to your Keep account.

There is also a Claude agent skill version of the same library access. Skill versus MCP is mostly a question of which client you are in and how chatty you want the tool integration to be. The library underneath is the same.

Where this sits next to Notion AI and Obsidian AI plugins

Notion AI does this for things you have written inside Notion. Obsidian's AI plugins do it for vaults you have built up in Obsidian. Both are good if your knowledge already lives there.

Keep's version is open across MCP-compatible clients. Your library is not trapped in a single app's chat panel. The same saves are queryable from Claude Code while you are deep in a codebase, from Claude Desktop while you are writing, and from Cursor while you are debugging. You pick the client. Keep just exposes the data.

If everything you save already lives in Notion or Obsidian and you are happy with the in-app AI, there is no reason to move. Keep earns its place when you save from many sources (browser extension, mobile share, email, X sync, RSS, YouTube, Kindle import) and want all of it queryable from whichever AI you happen to be using that hour.

A small habit that makes this much sharper

The MCP server is only as good as what is in the library. Save with the kind of prompt you might run against it later in mind. You do not need tags for every item. You do need to actually save the things you might want back, so a tab closing or a memory fading does not delete them.

I have asked Claude "find that thing I read about meditation and creativity" enough times to know the only failure mode is the one where I never saved it.

Read the MCP setup walkthrough for the full config across clients, or browse the agent skill if you would rather wire it up that way.